A Multi-label CNN model for the automatic detection and segmentation of primary brain tumors using 18F-FET PET imaging

Abstract Purpose The aim of this study was to develop a CNN model for automatic detection and segmentation of primary brain tumors using 18F-fluoroethyl-L-tyrosine (18F-FET) PET. Methods Eighty-four patients who underwent a 40 min dynamic 18F-FET PET were retrospectively included. Ground truth labels for the lesions and background regions were defined by a nuclear medicine physician using MIM software. A multi-label CNN was developed to segment the lesion and background region using the 20–40 min static scan. In addition, a single-label CNN was implemented for lesion-only segmentation. Lesion detectability was evaluated by classifying 18F-FET PET scans as negative when no tumor was segmented and vice versa, while the segmentation performance was assessed using the Dice similarity coefficient (DSC) and the segmented tumor volume. The quantitative accuracy was evaluated using the maximal tumor to average background uptake (TBR). Results The multi-label CNN model achieved a sensitivity of 88.9% and precision of 96.5% for discriminating between positive and negative 18F-FET PET scans which was significantly improved compared to the single-label CNN model with the sensitivity of 35.3% and precision of 83.1%. In addition, it allowed the accurate extraction of the maximal lesion and the average background uptake, resulting in an accurate TBR estimation compared to the fully manual approach. In terms of lesion detectability, the multi-label CNN model equally performed as the benchmark CNN model with the DSC of 74.5 ± 23.1 vs 73.7 ± 23.2. When using the tumor segmentations obtained by the two CNN models, the estimated tumor volume closely approximated the tumor volume estimated by the manual segmentations. Conclusion The proposed multi-label CNN model detected positive 18F-FET PET scans with high sensitivity and precision. Once detected, an accurate tumor segmentation and estimation of background activity was achieved resulting in an automatic, accurate TBR estimation, such that user interaction and potential inter-reader variability can be minimized.